# Native Platform (Linux / Windows / macOS)
## Runtime requirement
Ambi requires Tokio with multi-thread support. The minimal setup:
```toml
[dependencies]
tokio = { version = "1", features = ["rt-multi-thread", "sync", "time", "macros"] }
```
This is enforced by `Agent::make()` which calls `tokio::task::spawn_blocking()` to load the engine model without blocking the async runtime.
If you use `current_thread` runtime:
```rust
#[tokio::main(flavor = "current_thread")]
```
`Agent::make()` will panic because `spawn_blocking` requires a multi-thread runtime.
## `Agent::with_custom_engine` is different
`with_custom_engine()` wraps a `Box<dyn LLMEngineTrait>` directly and does **not** call `spawn_blocking`. It works with any Tokio runtime:
```rust
let agent = Agent::with_custom_engine(Box::new(MockEngine))?; // no spawn_blocking
```
## GPU acceleration
For llama.cpp local inference, GPU offloading is configured at build time via Cargo features:
```toml
# CUDA (NVIDIA)
ambi = { version = "0.3", features = ["llama-cpp", "cuda"] }
# Vulkan (multi-vendor)
ambi = { version = "0.3", features = ["llama-cpp", "vulkan"] }
# Metal (Apple Silicon)
ambi = { version = "0.3", features = ["llama-cpp", "metal"] }
# ROCm (AMD)
ambi = { version = "0.3", features = ["llama-cpp", "rocm"] }
```
Only one GPU backend can be enabled at compile time. Enabling two or more causes a `compile_error!`.
## Building from source
```bash
# Cloud only (fastest compile)
cargo build --no-default-features --features openai-api
# Local with CUDA
cargo build --features "llama-cpp, cuda"
```
## Known platform differences
- **Windows**: llama.cpp CUDA builds require the CUDA SDK and MSVC build tools. Use the `x64-native-nvidia` toolchain if available.
- **macOS**: Metal acceleration works on Apple Silicon (M1+). Intel Macs fall back to CPU.
- **Linux**: CUDA requires `libcuda.so` and NVCC in PATH. Vulkan needs the Vulkan SDK.